A speech recognizer optimally combining learning vector quantization, dynamic programming and multi-layer perceptron

Driancourt, X.; Gallinari, Patrick

doi:10.1109/icassp.1992.225835

Cited by 8 publications

(1 citation statement)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For example, a cost function based on a maximum mutual information criterion is used to train pure HMMs (Young 1991). Furthermore, it is shown in Driancourt and Gallinari (1992) how different kinds of ANN modules can be combined with DTW and trained using similar principles. This training could even be extended to DTW templates.…”

Section: F172 Speech Recognitionmentioning

confidence: 99%

Speech processing

Torkkola¹

Handbook of Neural Computation

View full text Add to dashboard Cite

Speech processing comprises automatic speech recognition, speech synthesis, speech coding, speech enhancement, speaker recognition and verification, language identification, and so on. This section discusses the application of artificial neural networks (ANNs) to these areas. The viewpoint will be that of an engineer; that is, the question we try to answer is, 'How can ANNs be used to solve engineering problems in speech processing?' We will present some conventional approaches to these problems and point out where ANNs could be applicable. As a lot of the ANN effort in speech processing seems to be concentrated around speech recognition, this will also be our focal point. Other areas will be briefly reviewed. Due to the breadth of the field and space limitations, this section can only remain superficial: more of a commented list of bibliographic references. F1.7.1 Introduction Speech is a medium for communication, and there is always a language behind it. While some speech processing applications can be regarded as pure 'signal processing', one usually cannot avoid taking into account that speech is a signal produced by human articulators. Furthermore, it may also be necessary to incorporate knowledge of the language to reach the best solutions. Thus, in addition to engineering, a successful speech processing application might need a combination of speech, hearing, and language sciences. Many kinds of characteristics of the speech signal can be learned by automated procedures using large databases. Statistical methods (including ANNs) rely on this fact. Some knowledge, especially linguistic, still needs to be obtained through manual coding and some needs to be taken into account implicitly. For instance, knowledge may be incorporated in the structure of a speech recognizer or a speech coder. Engineering problems in speech communication can roughly be placed in two categories: man-to-man communication and man-machine communication. Examples of the former category include speech coding for transmission and storage, speech enhancement, and speaker separation. Automatic speech recognition (ASR), speech synthesis, speaker identification and verification, and language identification would go in the latter category. Besides speech communication, another super-category is speech analysis, some aspects of which are necessary in every speech processing application. As here we can only touch the surface of these areas, the reader is encouraged to consult O'Shaughnessy (1987), Keller (1994), or Rabiner and Juang (1993) for background in speech processing. Some texts which concentrate on connectionist aspects of speech processing include Morgan and Scofield (1991), Bourlard and Morgan (1994), and Robinson (1993). Before going through the subcategories in detail, we will take a quick look at some of the generic capabilities of ANNs, and how they match problems in speech processing. It is well known that under some ideal conditions multilayer perceptrons (MLP) are universal approximators (Hornik et al 1989). There are C1.2 man...

show abstract

Section: F172 Speech Recognitionmentioning

confidence: 99%